Abliteration.ai offers access to AI models without guardrails
Abliteration.ai has turned the removal of refusal mechanisms from open-weight AI models into a commercial service. Its platform hosts modified systems, including Z.ai’s GLM-5.3, which customers can use through a browser or API. The company says the service is intended for offensive cybersecurity work, red-teaming and testing AI agents that standard models may refuse to support.
TechCrunch created a free account and found that the modified GLM-5.3 readily followed requests involving the theft of saved Chrome passwords and the home cultivation of a dangerous human pathogen. Abliteration.ai provides a moderation layer that customers can configure, while the platform itself retains some limited safeguards. Co-founder Devon said the company is working on additional protections, but it has not introduced identity verification beyond recording the payment card used by customers.
Safety researchers warn that wider access could facilitate real harm. Some cybersecurity companies say defenders need comparable tools because attackers may already use modified models, while others rely on fine-tuning and question whether abliteration preserves all of a model’s capabilities. The company has customers among European and UK red-teaming startups, has not raised venture capital and is discussing potential investment.
Written by Sakhanda Wire AI from reporting by TechCrunch. Facts, quotes and full context are in the original article below.